Transcript Intelligent Information Retrieval and Web Search

Probabilistic
Language-Model Based
Document Retrieval
1
Naïve Bayes for Retrieval
• Naïve Bayes can also be used for ad-hoc
document retrieval.
• Treat each of the n documents as a category with
only one training example, the document itself.
• Classify queries using this n-way categorization.
• Rank documents based on the posterior probability
of their category.
• For historical reasons, this is called the “language
model” (LM) approach.
Generative Model for Retrieval
Query
?
thisthis
testtest
urnurn
hashas
many
many
words
words
Doc1
0.3
this test
urn has
many
words
Doc2
0.1
?
?
thisthis
testtest
urnurn
hashas
many
many
words
words
Doc3
0.08
?
?
this test
urn has
many
words
Doc4
0.06
Ranked
Retrievals:
?
thisthis
testtest
urnurn
hashas
many
many
words
words
Doc5
0.18
this test
urn has
many
words
Doc6
0.28
Doc 1
Doc 6
Doc 5
Doc 2
Doc 3
Doc 4
0.3
0.28
0.18
0.1
0.08
0.06
Smoothing
• Proper smoothing is important for this
approach to work well.
• Laplace smoothing does not work well for
this application.
• Better to use linear interpolation for
smoothing.
Linear Interpolation Smoothing
• Estimate conditional probabilities P(Xi | Y) as a
mixture of conditioned and unconditioned
estimates:
P( X i | Y )  Pˆ ( X i | Y )  (1  )Pˆ ( X i )
^
• P(X
i | Y) is the probability of drawing word Xi from
the urn of words in category (i.e. document) Y.
^
• P(Xi) is the probability of drawing word Xi from the
urn of words in the entire corpus (i.e. all document
urns combined into one big urn).
Amount of Smoothing
• Value of λ controls the amount of smoothing.
• The lower λ is, the more smoothing there is since the
unconditioned term is weighted higher (1 ─ ).
• Setting λ properly is important for good performance.
• Set λ manually or automatically based on maximizing
performance on a development set of queries.
• Lower  tends to work better for long queries, high λ
for short queries.
Experimental Results on CF Corpus
Effect of  Parameter
Given long
queries, large
amount of
smoothing
(λ=0.1) seems
to work best
Experimental Results on CF Corpus
Effect of Laplace Smoothing Parameter
Experimental Results on CF Corpus
Comparison of Smothing Methods and VSR
Laplace smoothing
does much worse.
Linear interp does
about the same as
vector-space
Performance of Language Model Approach
• Larger scale TREC experiments
demonstrate that the LM approach with
proper smoothing works slightly better than
a well-tuned vector-space approach.
• Need to make LM approach efficient by
exploiting inverted index.
– Don’t bother to the compute probability of
documents that do not contain any of the query
words.